Directional Forecast Accuracy Across Hierarchical Aggregations Using Walmart M5 Data

This preregistration documents the planned analysis for a hierarchical forecast accuracy study using the Walmart M5 dataset. The study compares Bottom-Up (BU) and Top-Down (TD) forecasting structures, evaluating both directional accuracy (via binomial tests) and magnitude of improvement (via one-sample t-tests) across product and location hierarchies. MAE serves as the primary accuracy metric, with sMAPE and RMSE included for robustness. Forecasts are monthly, evaluated weekly, and use strictly pre-month historical data.

Deviation Note – Post-Processing Correction

During SQL import, some forecast tables contained negative forecast values, which are invalid for demand. 
A uniform correction was applied across all models and horizons: any negative forecast or lower-bound value 
was replaced with 0. Actuals were not changed. 

This deviation does not affect model fairness or comparative evaluation. All reported accuracy metrics are 
based on the corrected (clipped-at-zero) forecasts. 

Context Note (added November 2025):
This preregistration was originally motivated by practitioner claims that bottom-up forecasting improves accuracy by 1–3 % (SAP 2012; SAS 2014; Kinaxis 2016).
These figures are marketing estimates, not peer-reviewed findings.

The preregistered analysis plan itself remains unchanged; this note documents the study’s motivation and contextualizes the comparison discussed in the paper.